Apparatus and method for predicting tumor aggressiveness using a tumor- guided attention module

Abstract

The tumor malignancy prediction device and method using a tumor-induced attention module according to the embodiment propose a deep learning method for early prediction of the malignancy of prostate cancer, and enable effective learning by enhancing information on small tumors through the generation of size-normalized patches. Furthermore, the prostate cancer (PCa) invasiveness prediction model according to the embodiment first generates size-normalized patches to effectively capture the features of small tumors and the surrounding features of large tumors, and when size-normalized T2wMR, DWI, ADC maps, and tumor mask patches are generated according to the tumor size, they are used as inputs to the prostate cancer (PCa) malignancy prediction model. Subsequently, the embodiment uses a tumor-induced dual-branch network with multiple loss functions to better predict tumor invasiveness and distinguish between lPCa and hPCa.